Update app.py

app.py

@@ -65,19 +65,24 @@ def generate_atomic_facts(kb, question):
     Returns JSON: {"facts": [ ... ]}
     """
     prompt = f"""
-Extract atomic facts that directly answer the question.
+From the Knowledge Base, extract the character transformation of Matilda.
+
+Rules:
+- Identify INITIAL traits, CAUSAL EVENTS, and FINAL traits.
+- Use short factual statements grounded ONLY in the knowledge base.
+- Do NOT paraphrase the question.
+- Return facts that can be checked independently.
+
+Output strictly as JSON:
+{
+  "facts": [
+    "Initially Matilda desired a luxurious life despite her humble background",
+    "She pretended to be wealthy and borrowed a necklace to attend the ball",
+    "She lost the borrowed necklace, causing long-term suffering",
+    "As a result of hardship, she became mature, humble, and grateful"
+  ]
+}
 
-Knowledge Base:
-{kb}
-
-Question:
-{question}
-
-RULES:
-- Return 1-5 short factual statements that directly answer the question.
-- Output strictly in JSON format: {{"facts": ["fact1", "fact2", ...]}}
-- Do not include unrelated events or explanations.
-- Each fact should be self-contained.
 """
     inputs = llm_tokenizer(prompt, return_tensors="pt", truncation=True).to(DEVICE)
     outputs = llm_model.generate(
@@ -122,9 +127,17 @@ def evaluate_answer(answer, question, kb):
     covered_all = True
     for concept in schema["required_concepts"]:
         if claims:
-            scores = sim_model.predict([(concept, c) for c in claims])
-            best = float(scores.max())
-            ok = best >= SIM_THRESHOLD_REQUIRED
+
+            probs = softmax_logits(nli_model.predict([(c, concept)]))
+# index 2 = entailment for NLI DeBERTa
+            entailment = probs[2]
+            ok = entailment > 0.6
+            best = entailment
+
+
+            # scores = sim_model.predict([(concept, c) for c in claims])
+            # best = float(scores.max())
+            # ok = best >= SIM_THRESHOLD_REQUIRED
         else:
             best = 0.0
             ok = False
@@ -143,7 +156,13 @@ def evaluate_answer(answer, question, kb):
     for claim in claims:
         for sent in kb_sents:
             probs = softmax_logits(nli_model.predict([(sent, claim)]))
-            if probs[0] > CONTRADICTION_THRESHOLD:
+
+            contradiction = probs[0]
+            entailment = probs[2]
+
+            if contradiction > 0.8 and entailment < 0.2:
+            # probs = softmax_logits(nli_model.predict([(sent, claim)]))
+            # if probs[0] > CONTRADICTION_THRESHOLD:
                 contradictions.append({
                     "claim": claim,
                     "sentence": sent,